📚 New to this topic? Read our full guide: What Is Artificial Intelligence?.
Google Gemini 3.5 Flash Can Now See and Control Your Screen
Google has added a new power to its Gemini 3.5 Flash. Gemini 3.5 Flash is an AI model (a computer program that can think and answer in a human-like way). Now this model can look at a computer screen, see what is on it, and click, type, and scroll all by itself. Google calls this new skill “Computer Use.”
In simple words, the AI can now use a web browser, a phone app, or a computer the same way a person does. (A browser is the app you use to open websites, like Chrome.) This makes Gemini 3.5 Flash an “agentic” model. Agentic means the AI does not just answer your questions. It also does the work for you, one step at a time.
Until now, this screen-control skill lived in a different, older model called Gemini 2.5. Putting it right inside Gemini 3.5 Flash is a big deal. That is because Flash is Google’s fast and cheap model. So coders get screen control without paying for a bigger, costlier model.
What “computer control” actually means
Most AI chatbots can only talk. This one can also act. You give Gemini 3.5 Flash a goal. Then it looks at the screen, decides where to click, and does the job. Google says the model can “see, understand, and interact with computers, browsers, and mobile devices on its own.”
This opens up real work. A developer (a person who builds apps and websites) could ask the model to test a website by clicking every button. A business could ask it to fill in forms or move data from one app to another. The model works on browsers, phones, and computers.
How good is it? The benchmark scores
Google tested the model using OSWorld. OSWorld is a benchmark, which is a standard test used to compare AI models. This test checks how well an AI can finish real jobs on a computer desktop. A higher score is better. Here is how Gemini 3.5 Flash did against other models on the reported OSWorld scores.
| Model | OSWorld score (computer control) |
|---|---|
| Anthropic Opus 4.8 | 83.4 |
| GPT-5.5 | 78.7 |
| Gemini 3.5 Flash | 78.4 |
| Sonnet 4.6 | 78.4 |
| Gemini 3.1 Pro | 76.2 |
| GPT-5.4 mini | 72.1 |
| Gemini 3 Flash | 65.1 |
Benchmarks & specs at a glance
| Spec | Detail (as reported) |
|---|---|
| Model | Gemini 3.5 Flash |
| New skill | Computer Use (screen control) built in |
| OSWorld score | 78.4 |
| Works on | Browser, mobile, desktop |
| Where to get it | Gemini API, Gemini Enterprise Agent Platform |
| Extras | Browserbase demo, GitHub reference code |
What about safety?
Letting an AI click on its own can be risky. So Google added some guards. The company used “adversarial training.” This means they trained the model with tricky and harmful inputs on purpose, so it learns to handle them. Google also gave businesses two safety switches they can turn on.
The first switch asks a real person to say yes before the AI does anything risky, like making a payment. The second switch stops the task on its own if the model spots a “prompt injection.” A prompt injection is a hidden trick on a webpage that tries to fool the AI into doing the wrong thing. Google also tells companies to run the AI inside a sandbox. (A sandbox is a safe, closed space where the AI can act without harming real systems.) And it says a human should always keep watch.
Where you can use it
Developers can use the feature through the Gemini API and the Gemini Enterprise Agent Platform. (An API is a way for one program to talk to another and use its features. “Enterprise” just means it is made for big companies.) Google also shared a live Browserbase demo and sample code on GitHub, so coders can start quickly. This launch is part of a bigger push by Google into AI agents for businesses. It is similar to how Indian IT firms are now building AI agents on Gemini Enterprise.
FAQ
What is Gemini 3.5 Flash?
It is Google’s fast, low-cost AI model. The new version can also see and control a computer screen on its own.
What is OSWorld?
It is a test that scores how well an AI finishes real tasks on a desktop. Gemini 3.5 Flash scored 78.4 out of 100.
Is it safe to let AI control my screen?
Google added safety switches, like asking a human to confirm risky actions and stopping if it spots a hidden trick. But experts still suggest running it in a controlled space with human checks.
Why it matters (especially for India and founders)
India runs a lot on services and software work. A fast, cheap AI that can use apps could change how startups build their products. A small team could let the AI do software testing, data entry, or customer tasks. That means they may not need to hire many people. Because Flash is cheap, even bootstrapped Indian founders can try it. (Bootstrapped means they run on their own money, without big outside funding.) This is also part of a bigger race, with India leading global demand for AI agents and assistants. The companies that learn to use screen-controlling AI early could move much faster than their rivals.
The main idea is simple. AI is moving from just talking to actually doing. Gemini 3.5 Flash can now see your screen and act on it, at a low price. The big question for every founder is no longer “can AI answer this?” It is now “what work can I hand over to an AI agent?”
Source: The Decoder.
Related coverage
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
